fix(mm): identify Qwen2.5/Qwen3 causal LMs as Text LLM, not Qwen3 encoder

[Pfannkuchensack]

Summary

Qwen3Encoder_Qwen3Encoder_Config matches any directory with config.json at the root and a Qwen* class name, which also describes a complete causal LM like Qwen2.5-1.5B-Instruct. With both configs matching and equal sort keys, the encoder won the tie-break, blocking the model from being used as a prompt expander.

A standalone text_encoder download has no tokenizer files; a complete causal LM does. Use that to reject the encoder match when tokenizer files are present alongside config.json at the root.

Also fixes the secondary failure when manually switching the type from Qwen3Encoder to Text LLM: the existing record's format=qwen3_encoder and variant were carried over and produced no matching discriminator under text_llm. On ValidationError after a type change, retry with the stale fields stripped so the new class can apply its defaults.

Closes #9090

Related Issues / Discussions

• Closes [bug]: Qwen2.5-1.5B-Instruct model not recognized as text_llm #9090
• Originally introduced via the prompt-expander work in Feat(UI): Add LLM-powered prompt expansion and image-to-prompt features #8899

QA Instructions

Primary bug — probe misidentification:

1. Open Model Manager → Add Model → HuggingFace.
2. Install Qwen/Qwen2.5-1.5B-Instruct.
3. Verify the model is identified as Text LLM (not Qwen3 Encoder) and is usable as a prompt expander.

Secondary bug — manual type switch:

1. On main (without this PR), install Qwen2.5-1.5B-Instruct; it lands as Qwen3 Encoder.
2. With this PR applied, edit the existing record and change the type to Text LLM → save.
3. Expected: save succeeds and the record is now a TextLLM_Diffusers_Config. Previously this returned "Model Update Failed".

Regression checks:

• Install a real Qwen3 encoder (e.g. the text_encoder subfolder bundled with a Z-Image / Qwen-Image pipeline) — should still be identified as Qwen3 Encoder.
• Install a full diffusers pipeline that contains a text_encoder/ subfolder — unchanged behavior.
• Run pytest tests/app/services/model_records/ tests/backend/model_manager/configs/ → 31 tests pass.

Merge Plan

Standard merge — no schema/migration changes, no sensitive surfaces touched. Backend-only fix.

Checklist

• The PR has a short but descriptive title, suitable for a changelog
• Tests added / updated (if applicable)
• ❗Changes to a redux slice have a corresponding migration
• Documentation added / updated (if applicable)
• Updated What's New copy (if doing a release after this PR)

[lstein]

Works as advertised.

[lstein]

Identification and reidentification are working properly. Just a note on the secondary bug, that when I install Qwen2.5-1.5B-Instruct before applying the PR, apply the PR, and then try to change its type without also changing its format, I get this error:

[2026-05-04 21:44:27,107]::[InvokeAI]::ERROR --> 1 validation error for TextLLM_Diffusers_Config
format
  Input should be <ModelFormat.Diffusers: 'diffusers'> [type=literal_error, input_value='qwen3_encoder', input_type=str]
    For further information visit https://errors.pydantic.dev/2.12/v/literal_error

I have to change the Model Format to diffusers as well as changing the Model Type, after which all is well. Please let me know if this is the expected behavior before I merge this.

[Pfannkuchensack]

Yes, the Format of that Model is diffusers.

[Pfannkuchensack]